Measuring operational risk in financial institutions: …...1 Measuring operational risk in...
Transcript of Measuring operational risk in financial institutions: …...1 Measuring operational risk in...
1
Measuring operational risk in financial institutions:
Contribution of credit risk modelling
HÜBNER Georges, PhD1
PETERS Jean-Philippe2
PLUNUS Séverine3
Key words: Operational risk, modelling, CreditRisk+.
1 The Deloitte Professor of Financial Management at HEC-University of Liège, Associate Professor of Finance
at Maastricht University, and Senior Researcher at the Luxembourg School of Finance, University of
Luxembourg. 2 Deloitte Luxembourg, Advisory and Consulting Group (Risk Management Unit), . Ph.D. Candidate, HEC-
Management School, Université de Liège, Belgium E-mail: [email protected]. 3 Corresponding Author. Ph.D. Candidate, HEC-Management School,University of Liège. Rue Louvrex, 14, B-
4000 Liège, Belgium. E-mail: [email protected]. Tel. (+32)42327429.
2
ABSTRACT
The scarcity of internal loss databases tends to hinder the use of the advanced approaches for
operational risk measurement (AMA) in financial institutions. As there is a greater variety in credit
risk modelling, this paper explores the applicability of a modified version of CreditRisk+ to
operational loss data. Our adapted model, OpRisk+, works out very satisfying Values-at-Risk at 95%
level as compared with estimates drawn from sophisticated AMA models. OpRisk+ proves to be
especially worthy in the case of small samples, where more complex methods cannot be applied.
OpRisk+ could therefore be used to fit the body of the distribution of operational losses up to the
95%-percentile, while Extreme Value Theory or external databases should be used beyond this
quantile.
Key words: Operational risk, modelling, CreditRisk+.
3
1. Introduction
Over the past decade, financial institutions have experienced several large operational loss events
leading to big banking failures. Memorable examples include the Barings’ bankruptcy in 1995, the
$691 million trading loss at Allfirst Financial, or the $140 million loss at the Bank of New York due
to September 11th, 2001. These events, as well as developments such as the growth of e-commerce,
changes in banks’ risks management or the use of more highly automated technology, have led
regulators and the banking industry to recognize the importance of operational risk in shaping the risk
profiles of financial institutions.
Reflecting this recognition, the Basel Committee on Banking Supervision, in its proposal for A New
Capital Accord, has incorporated into its proposed capital framework an explicit capital requirement
for operational risk, defined as the risk of loss resulting from inadequate or failed internal processes,
people and systems or from external events. This definition includes legal risk, but not strategic and
reputational risks. As for credit risk, the Basel Committee does not believe in a “one-size-fits-all”
approach to capital adequacy and proposes three distinct options for the calculation of the capital
charge for operational risk: the basic indicator approach, the standardized approach and the advanced
measurement approaches (AMA). The use of these approaches of increasing risk sensitivity is
determined according to the risk management systems of the banks. The first two methods are a
function of gross income, while the advanced methods are based on internal loss data, external loss
data, scenario analysis, business environment and internal control factors.
In 2001, the Basel Committee was encouraging two specific AMA methods: (i) the Loss Distribution
Approach (LDA) and (ii) an Internal Measurement Approach (IMA) developing a linear relationship
between unexpected loss and expected loss to extrapolate credit-risk’s internal rating based (IRB)
approach to operational risk. However, prior to its 2003 Accord, the Basel Committee dropped formal
mention of the IMA leaving the LDA as the only specifically recommended AMA method.
Even though the Basel Accord formally dropped Internal Measurement Approaches in favour of
Value-at-Risk approaches, it is still legitimate to be inspired by modelling approaches for credit risk in
order to model the distribution of operational loss data. Indeed, both definitions of risks have similar
features, such as their focus on a one-year measurement horizon or their use of an aggregated loss
distribution skewed towards zero with a long right-tail.
This paper explores the possibility of adapting one of the current proposed industry credit-risk models
to perform much of the functionality of an actuarial LDA model. We identified CreditRisk+, the
model developed by Credit Suisse, as an actuarial-based model whose characteristics can be adapted
to fit the Loss Distribution Approach, which is explicitly mentioned in the Basel II Accord as eligible
4
among the Advanced Measurement Approaches (AMA) to estimate risk capital, and has
unambiguously emerged as the standard industry practice (see Sahay et al. (2007) or Degen et al.
(2007)) . After some adjustment, we construct a distribution of operational losses through the adapted
“OpRisk+” model. As this model calibrates the whole distribution, not only can we retrieve the
quantiles of the operational loss distribution, but also an estimate of its expectation, needed for the
computation of the economic capital.
Our research is aimed at answering the following question: how would the adaptation of CreditRisk+
model perform compared to sophisticated models such as the approach developed by Chapelle,
Crama, Hübner and Peters (2008) (henceforth CCHP) or Moscadelli (2004) among others or
compared to an extended IMA approach such as Alexander (2003)?
We address the question with an experiment based on generated databases using three different Pareto
distributions. We study the behaviour of OpRisk+, together with an alternative characterization that
specifically aims at modelling operational losses, when all these models are confronted to either fat,
medium or thin tails for the loss distributions. Furthermore, we assess the influence of the number of
losses recorded in the database on the quality of the estimation. The knowledge of the true distribution
of losses is necessary to assess the quality of the different fitting methods. Had a real data set been
used instead of controlled numerical simulations, we would not be able to benchmark the observed
results against the true loss distribution and therefore we could not assess the performance of OpRisk+
for different loss generating processes and sample sizes.
We also test our new adapted IRB model against Alexander’s existing “lower-bound” improvement to
the basic IMA formula to see if our model outperforms. The lower bound is effectively a quantile
value from a normal distribution table which allows identification of the unexpected loss if you know
the mean and variance of the loss severity distribution and the mean of the frequency distribution. The
multiplier is called the “lower-bound” because it will have its lowest value (approximately equal to
3.1) when the mean of the frequency distribution corresponds to a high-frequency type of loss (e.g.
more than 100 losses per year).
Our main findings are twofold. First, we note that the precision of OpRisk+ is not satisfactory to
estimate the very far end of the loss distribution, such as the Value-at-Risk (VaR)1 at the 99.9%
confidence level. Yet, our model works out very satisfying quantile estimates, especially for thin-
tailed Pareto-distributions, up to a 95% confidence level for the computation of the VaR. The
estimation error is relatively small and stable across distributions. Secondly, the simplicity of our
model makes it applicable to “problematic” business lines, that is, with very few occurrences of
5
events, and with very few years of data. Procedures that rely on extreme-value theory, by contrast, are
very data-consuming, and yield very poor results when used with small databases.
These findings make the OpRisk+ approach clearly not an effective substitute, but indeed a very
useful complement to approaches that specifically target the extreme tail of the loss distribution. In
particular, the body of the loss distribution can be safely assessed with our method, while external
data, as specifically mentioned in the Accord, can be used to estimate the tail. Being able to
simultaneously rely on the body and the tail of the distribution is crucial for the operation risk capital
estimation, because one needs the full distribution of losses in order to capture the expected loss that
enters the regulatory capital estimate.
The paper is organized as follows: Section 2 describes the adjustment needed in order to apply
CreditRisk+ model to operational loss data and presents two alternative methods to calibrate a Value-
at-Risk on operational loss data (OpVaR). Section 3 describes our database, presents our results and
compares them to the other approaches’ results. Section 4 concludes.
2. Alternative Approaches for the Measurement of Operational Risk
This section first presents three alternative ways to calibrate a Value-at-Risk on operational loss data.
The first one represents an adaptation of the CreditRisk+ framework, while the second one proposes
an adaptation of the Loss Distribution Approach (LDA) in the context of operational losses with the
use of Extreme Value Theory (EVT). Finally, we introduce a lower bound for the Value-at-Risk
derived from a model developed by Alexander (2003).
2.1 OpRisk+: Application of CreditRisk+ to Operational Loss Data
CreditRisk+ developed by Credit Suisse First Boston is an actuarial model derived from insurance
losses models. It models the default risk of a bond portfolio through the Poisson distribution. Its basic
building block is simply the probability of default of a counterparty. In this model, no assumptions are
made about the causes of default: an obligor is either in default with a probability PA, or not in default
with a probability 1-PA. Although operational losses do not depend on a particular counterparty, this
characteristic already simplifies the adaptation of our model, as we do not need to make assumptions
on the causes of the loss.
6
CreditRisk+ determines the distribution of default losses in three steps: the determination of the
frequency of defaults, approximated by a standard Poisson distribution, the determination of the
severity of the losses and the determination of the distribution of default losses.
The determination of the frequency of events leading to operational losses can be modelled through
the Poisson distribution as for the probability of default in CreditRisk+:
(1)
where μ is the average number of defaults per period, and N is a stochastic variable with mean µ, and
standard deviation √μ.
CreditRisk+ computes the parameter μ by adding the probability of default of each obligor, supplied,
for instance, by rating agencies. However, operational losses do not depend on a particular obligor.
Therefore, instead of being defined as a sum of probabilities of default depending on the
characteristics of a counterpart, µ can be interpreted as the average number of loss events of one type
occurring in a specific business line during one period.
CreditRisk+ adds the assumption that the mean default rate is itself stochastic in order to take into
account the fat right tail of the distribution of defaults. Nevertheless, CCHP (2008) show that, as far as
operational risk is concerned, using the Poisson distribution to model the frequency of operational
losses does not lead to a substantial measurement error. Hence, we keep on assuming that the number
of operational loss events follows a Poisson distribution with a fixed mean µ.
In order to perform its calculations, CreditRisk+ proposes to express the exposure (here, the losses) in
a unit amount of exposure L2. The key step is then to round up each exposure size to the nearest
whole number, in order to reduce the number of possible values and to distribute them into different
bands. Each band is characterized by an average exposure, νj and an expected loss, εj, equal to the
sum of the expected losses of all the obligors belonging to the band. Table 1 shows an example of this
procedure.
[TABLE 1]
CreditRisk+ posits that
(2)
where εj is the expected loss in band j, νj is the common exposure in band j, and μj is the expected
number of defaults in band j.
7
As the operational losses do not depend on a particular transaction, we slightly modify the definition
of these variables. The aim is to calculate the expected aggregate loss. We will therefore keep the
definition of εj unchanged. However, as noted earlier, μj is not an aggregate expected number of
defaults anymore but simply the (observed) average number of operational loss events occurring in
one year. Consequently, in order to satisfy equation (2), νj must be defined as the average loss amount
per event for band j. The following table illustrates the reprocessing of the data:
[TABLE 2]
Each band is viewed as a portfolio of exposures by itself. Because some defaults lead to larger losses
than others through the variation in exposure amounts, the loss given default involves a second
element of randomness, which is mathematically described through its probability generating
function. Thus, let G(z) be the probability generating function for losses expressed in multiples of the
unit L of exposure:
(3)
As the number of defaults follows a Poisson distribution, this is equal to:
(4)
As far as operational losses are concerned, a band can no more be considered as a portfolio but will
simply be seen as a category of loss size. This also simplifies the model, as we do not distinguish
exposure and expected loss anymore. For credit losses, exposures are first sorted, and then the
expected loss is calculated, by multiplying the exposures by their probability of default. As far as
operational losses are concerned, the loss amounts are directly sorted by size. Consequently, the
second element of randomness is not necessary anymore. This has no consequences on the following
results except simplifying the model.
Whereas CreditRisk+ assumed the exposures in the portfolio to be independent, OpRisk+ will assume
the independence of the different loss amounts. Thanks to this assumption, the probability generating
function for losses of one type for a specific business line is given by the product of the probability
generating function for each band:
8
(5)
Finally, the loss distribution of the entire portfolio is given by:
(6)
Note that this equation allows only computing the probability of losses and size 0, L, 2L and so on.
This probability of loss of nL will further be denoted An.
Then, under the simplified assumption of fixed default rates, Credit Suisse has developed the
following recursive equation3:
(7)
where
m
j j
j
eeGA1
)0(0 .
The calculation depends only on 2 sets of parameters: ν j and εj, derived from μj , ,the number of events
of each range, j, observed. With operational data, A0 is derived directly from eA0 .
To illustrate this recurrence, suppose your database contains 20 losses, 3 (resp. 2) of which having a
size of 1L (resp. 2L):
9200 10.06.2eA
96.18.1092.06.10 x 31
011
1:
1 AAAj
jj
j
Therefore, the probability of having a loss of size resp. 0, 1L and 2L is resp. 2.06.10-9
, 6.18.10-9 and
1.13.10-8
, and so on.
From there, you can re-construct the distribution of the loss of size nL.
The procedure can be summarized as follows:
1. Choose a unit amount of loss L4.
2. Divide the losses of the available database by L and round up these numbers.
9
3. Allocate the losses of different sizes to their band and compute the expected loss per
band, equal to the observed number of losses per band multiplied by the average loss amount
per band, equal to j.
4. Compute the probability of zero losses equal to A0 = e-µ
, where µ is the total number of
losses observed per period.
5. For each band j = 1 to n, compute the probability of losses of size n by using equation
(7).
6. Cumulate the probabilities in order to find the OpVaR99.9.
7. Repeat the procedure for each year of data.
2.2 The Loss Distribution Approach adapted to Operational Risk
Among the Advanced Measurement Approaches (AMA) developed over the recent years to model
operational risk, the most common one is the Loss Distribution Approach (LDA), which is derived
from actuarial techniques (see Frachot, Georges and Roncalli, 2001 for an introduction).
By means of convolution, this technique derives the aggregated loss distribution (ALD) through the
combination of the frequency distribution of loss events and the severity distribution of a loss given
event5. The operational Value-at-Risk is then simply the 99.9
th percentile of the ALD. As an analytical
solution is very difficult to compute with this type of convolution, Monte Carlo simulations are
usually used to do the job. Using the CCHP procedure with a Poisson distribution with a parameter µ
equal to the number of observed losses during the whole period to model the frequency6, we have:
1. Generate a large number M of Poisson(µ) random variables (say, 10000). These M values
represent the number of events for each of the M simulated periods.
2. For each period, generate the required number of severity random variables (that is, if the
simulated number of events for period m is x, then simulate x severity losses) and add
them to get the aggregated loss for the period.
3. The obtained vector represents M simulated periods. If M = 10000, when sorted, the
smallest value thus represents the 0.0001 quantile, the second the 0.0002 quantile, etc.,
which makes the OpVaRs very easy to calculate.
Examples of empirical studies using this technique for operational risk include Moscadelli (2004) on
loss data collected from the Quantitative Impact Study (QIS) of the Basel Committee, de Fontnouvelle
and Rosengren (2004) on loss data from the 2002 Risk Loss Data Collection Exercise initiated by the
Risk Management Group of the Basel Committee or CCHP with loss data coming from a large
European bank.
10
The latter underline that, in front of an actual database of internal operational losses, mixing two
distributions fit more adequately the empirical severity distribution than a single distribution.
Therefore, they divide the sample into two parts: a first one with losses below a selected threshold,
considered as the “normal” losses, and a second one, including the “large” losses. To model the
“normal” losses, CCHP compare several classic continuous distributions such as gamma, lognormal or
Pareto.
To take extreme and very rare losses into account (i.e. the “large” losses), the authors apply the
Extreme Value Theory (EVT) on their results7. The advantage of EVT is that it provides a tool to
estimate rare and not-yet-recorded events for a given database8. The use of EVT works out very
similar OpVaRs, except for the very high percentile, that is the 99.99th
. As the Basel Accord requires a
confidence level of 99.90, the authors obtain that their Monte Carlo simulation allows them to
compute a sufficiently complete sample, including most of the extreme cases.
2.3 Comparison with a lower bound
The basic formula of the Internal Measurement Approach (IMA) included in the Advanced
Measurement Approaches of Basel II is:
ELUL (8)
where UL = unexpected loss, determining the operational risk requirement9, and is a multiplier.
Gamma factors are not easy to evaluate as no indication of their possible range has been given by the
Basel Committee. Therefore, Alexander (2003) suggests that instead of writing the unexpected loss as
a multiple (γ) of expected loss, one writes unexpected loss as a multiple (Φ) of the loss standard
deviation. Using the definition of the expected loss, she gets the expression for Φ:
ELaR 9.99V (9)
The advantages of this parameter are that it can be easily calibrated and that it has a lower bound.
The basic IMA formula is based on the binomial loss frequency distribution, with no variability in loss
severity. For very high-frequency risks, Alexander notes that the normal distribution could be used as
an approximation of the binomial loss distribution, providing for Φ a lower bound equal to 3.1 (as can
be found from standard normal tables when the number of losses goes to infinity). She also suggests
that the Poisson distribution should be preferred to the binomial as the number of transactions is
generally difficult to quantify.
11
Alexander (2003) also shows that Φ, as a function of the parameter of the Poisson distribution,
must be in a fairly narrow range: from about 3.2 for medium-to high frequency risks (20 to 100 loss
events per year) to about 3.9 for low frequency risks (one loss event every one or two years) and only
above 4 for very rare events that may happen only once every five years or so. Table 3 illustrates the
wide range for the gammas by opposition to the narrow range of the phi’s values.
[TABLE 3]
Then, assuming the loss severity to be random, i.e. with mean µL and standard deviation σL, and
independent of the loss frequency, Alexander writes the Φ parameter as:
(10)
Where λ is the average number of losses.
For σL > 0, this formula produces slightly lower Φ than with no severity uncertainty, but it is still
bounded below by the value 3.1.
For a comparison purpose, we will use the following value for the lower bound of the needed
OpVaR99.9, derived from equation (10) in which we replaced Φ by 3.1 (asymptotic value for the ratio
of the difference between the 99th quantile and the parameter of the Poisson distribution to its standard
deviation) :
(11)
Our model should therefore produce operational economic capital at least equal or higher than this
lower bound.
3. An Experiment on Simulated Losses
3.1 Data
OpRisk+ makes the traditional statistical tests impossible, as it uses no parametrical form but a purely
numerical procedure. Therefore, in order to perform tests of the calibrating performance of OpRisk+
on any distribution of loss severity, we simulate databases to obtain an exhaustive picture of the
capabilities of the approach, that is, on the basis of three different kinds of distributions: a heavy-tail, a
medium-tail and a thin-tail Pareto distribution. Indeed, Moscadelli (2004) and de Fontnouvelle and
12
Rosengren (2004) have shown that loss data for most business lines and event types may be well
modelled by a Pareto-type distribution.
A Pareto distribution is a right-skewed distribution parameterized by two quantities: a minimum
possible value or location parameter, xm, and a tail index or shape parameter, ξ. Therefore, if X is a
random variable with a Pareto distribution, the probability that X is greater than some number x is
given by:
for all x ≥ xm., and for xm and k= 1/ξ >0.
The parameters of our distributions are Pareto(100;0.3), Pareto(100;0.5) and Pareto(100;0.7): the
larger the value of the tail index, the fatter the tail of the distribution. The choice of these functions has
been found to be reasonable with a sample of real data obtained from a large European institution.
We ran three simulations: one for the thin-tailed Pareto severity distribution case, one for the medium-
tailed Pareto severity distribution case and one for the fat-tailed Pareto severity distribution case. For
each of these cases, we simulated two sets of 1000 years of 20 and respectively 50 operational losses
and two sets of 10010
series of 200 losses and 300, respectively. For each of the 6600 simulated years
(3 x 2 x 1100), the aggregated loss distribution has been computed with the algorithm described in
Section 2.2.
Table 4 gives the characteristics of each of the twelve databases (each thus comprising 1000 or 100
simulated aggregate loss distributions) constructed in order to implement OpRisk+. For each series of
operational losses we computed the expected loss, that is, the mean loss multiplied by the number of
losses. The five last lines present the mean, standard deviation, median, maximum and minimum of
these expected losses.
[TABLE 4]
These results clearly show that data generated with a thin-tailed Pareto-distribution exhibit
characteristics that make the samples quite reliable. The mean loss is very close to its theoretical level
even for 20 draws. Furthermore, we observe a standard deviation of aggregate loss that is very limited,
from less than 10% of the average for N=20 to less than 3% for N=200. The median loss is also close
to the theoretical value. For a tail index of 0.5 (medium-tailed), the mean loss still stays close to the
13
theoretical value but the standard deviation increases. Thus, we can start to question the stability of the
loss estimate.
When the tail index increases, all these nice properties collapse. The mean aggregate loss becomes
systematically lower than the theoretical mean, and this effect aggravates when one takes a lower
number of simulations (100 drawings) with a larger sample. The standard deviation and range become
extremely large, making inference based on a given set of loss observations extremely adventurous.
3.2. Application of OpRisk+
To apply OpRisk+ to these data, the first step consists of computing A0 = e-μ
, where μ is the average
number of loss events. For instance, for N=200, this gives the following value:
87200
0 1038.1eA . Then, in order to assess the loss distribution of the entire population of
operational risk events, we use the recursive equation (7) to compute A1, A2 etc.
Once the different probabilities An for the different sizes of losses are computed, we can plot the
aggregated loss distribution as illustrated in Figure 1.
[FIGURE 1]
With this information, we can compute the different Operational Values-at-Risk (OpVaR). This is
done by calculating the cumulated probabilities for each amount of loss. The loss for which the
cumulated probability is equal to p% gives us the OpVaR at percentile p.
The average values for the different OpVaRs are given in Tables 4 and 5. Table 4 compares the
OpVaRs obtained using OpRisk+ with the simulated data for the small databases. The first column
represent the average observed quantiles of the aggregated distribution when simulating 25000 years
with a Poisson(mu) distribution for the frequency and a Pareto(100, ξ.) for the severity. The tables also
gives the minimum, maximum and standard deviation of the 100(0) OpVaRs produced by OpRisk+.
Panel A of Table 5 shows that OpRisk+ achieves very satisfactory OpVaRs for the Pareto-distribution
with thin tail. The mean OpVaRs obtained for both the samples of 20 and 50 observations stays within
a 4% distance from the true value. Even at the level of 99.9% required by Basel II, the OpRisk+
14
values stay within a very narrow range, while the standard deviation of the estimates is kept within
10% around the good value.
[TABLE 5]
The results obtained with the OpRisk+ procedure with medium and fat tails tend to deteriorate, which
is actually not surprising as the adaptation of the credit risk model strictly uses observed data and does
necessarily underestimate the fatness of the tails. However, we still have very good estimation for
OpVaR95. It mismatches the true 95% quantile by 2% to 7% for the medium and fat tailed Pareto-
distribution, while the standard deviation tends – naturally – to increase very fast.
The bad news is that the procedure alone is not sufficient to provide the OpVaR99.9 required by Basel
II. It severely underestimates the true quantile, even though this true value is included in the range of
the observed values of the loss estimates.
Table 6 displays the results of the simulations when a large sample size is used.
[TABLE 6]
Table 6, Panel A already delivers some rather surprising results. The OpRisk+ procedure seems to
overestimates the true operational risk exposure for all confidence levels; this effect aggravates for a
high number of losses in the database. This phenomenon is probably due to an intervalling effect,
where losses belonging to a given band are given the value of the band. Given that extreme losses are
likely to occur in the lower part of the band, as the distribution is characterized by a thin tail Pareto-
distribution, taking the upper bound limit value for aggregation seems to deteriorate the estimation,
making it too conservative. Nevertheless, the bias is almost constant in relative terms, indicating that
its seriousness does not aggravate as the estimation gets far in the tail of the distribution. Sub-section
5.4. will go further in this assumption.
This intervalling phenomenon explains the behaviour of the estimation for larger values of the tail
index. In Panel B, the adapted credit risk model still overestimates the distribution of losses up to a
confidence level of 99%, but then does not capture to distribution at the extreme end of the tail
15
(99.9%). In Panel C, the underestimation starts earlier, around the 90% percentile of the distribution.
The procedure, using only observed data, is still totally unable to capture the fat tailedness of the
distribution of aggregated losses.
Nevertheless, from panels B and C altogether, the performance of OpRisk+ still stays honourable
when the confidence level of 95% is adopted. The standard deviation of the estimates also remains
within 15% (with the tail index of 0.5) and 30% of the mean (with a tail index of 0.7), which is fairly
large but mostly driven by large outliers as witnessed in the last column of each panel.
A correct mean estimate of the OpVaR95 would apply to a tail index between 0.5 and 0.7, which
corresponds to a distribution with a fairly large tail index. Only when the tail of the Pareto-distribution
is actually thin, one observes that the intervalling effect induces a large discrepancy between the
theoretical and observed values.
It remains to be mentioned that the good application of OpRisk+ does not depend on the number of
observed losses as it only affects the first term of the recurrence, namely A0.
3.3. Comparison with the CCHP approach and the Lower-Bound of Alexander.
These results, if their economic and statistical significance have to be assessed, have to be compared
with a method that aims at specifically addressing the issue of operational losses in the Advanced
Measurement Approaches setup. We choose the CCHP approach, which is by definition more
sensitive to extreme events than OpRisk+, but has the drawback of requiring a large number of events
to properly derive the severity distributions of “normal” and “large” losses.
The graphs from Figure 2 display the OpVaRs (with confidence levels of 90, 95, 99 and 99.9%)
generated from three different kind of approaches, that is the sophisticated CCHP approach, OpRisk+
and the simpler Alexander (2003) approach (See Section 4.3) for three small databases of 20 and of 50
loss events, and for three larger databases of 200 and of 300 events.
[FIGURE 2]
From the graphs in Figure 2, we can see that for most databases, OpRisk+ is working out a capital
requirement higher than the lower bound, but smaller than the CCHP approach. This last result could
be expected as CCHP is more sensitive to extreme events. We will discuss the fact that the database
16
with 300 observations shows higher OpVaRs for OpRisk+ than CCHP in the last sub-section.
However, we can already conclude that our model is more risk sensitive than a simple IMA approach.
Considering the thin tailed Pareto-distribution in Panel A, we can observe that OpRisk+ produces the
best estimations for the small database. Indeed, those are very close to the theoretical OpVaRs for all
confidence level. However, for the large database, it is producing too cautious (large) OpVaRs. The
comparison with other methods sheds new light on the results obtained with Panel A of Table 6:
OpRisk+ overestimates the true VaR, but the CCHP model, especially dedicated to the measurement
of operational risk, does even worse. Actually, Alexander’s (2003) lower bound, also using observed
data but not suffering from an intervalling effect, works out very satisfactory results when the standard
deviation of loss is a good proxy of the variability of the distribution.
For the medium and fat tailed Pareto-distributions, neither of the models is sensitive enough for
OpVaRs of 99% and more. However, as far as the small databases are concerned, it is interesting to
note that OpRisk+ is producing the best estimations for OpVaR95.
Figure 3 compares the different values for the OpVaR90, OpVaR95, OpVaR99 and OpVaR99.9, for one of
the small samples and one of the large samples. It shows that the CCHP model produces a heavier tail
than OpRisk+ does, except for the Pareto(200;0,7), where they are quite similar. However, OpRisk+
yields OpVaRs that are much closer to the theoretical ones as far as small databases are concerned. It
even consistently produces very good OpVaR95 in all cases. However, at the level of confidence
required by Basel II, this procedure leaves estimates that are still far from the expected value,
especially when the tail of the Pareto-distribution gets bigger.
[ FIGURE 3]
3.4. Comparison with OpRisk+ taking an average value of loss for each band.
As shown above, taking the upper bound limit value for aggregation as described in the CreditRisk+
model tends to overestimate the true operational risk exposure for all confidence levels; especially
with larger databases. A solution could be to take the average value of losses for each band11
. Table 7
displays the results of the simulations when a relatively large sample size is used.
Panel A of Table 7 shows that OpRisk+ achieves very good results for the Pareto-distribution
characterized by a thin tail when using an average value for each band (“round” column). The OpVaR
values obtained for the sample of 200 observations is very close from the theoretical value, whereas it
stays within a 6% range from the “true” value with a 300 observations sample, including at the Basel
II level of 99.9%.
17
When the loss Pareto-distributions are medium-tailed, the results obtained with the OpRisk+
procedure with the databases are very good for quantiles up to 95% but deteriorate for more sensitive
OpVaRs. OpRisk+ is still totally unable to capture the tailedness of the distribution of aggregated
losses for very high confidence interval, such as the Basel II requirement.
[TABLE 7]
Table 8 compares the two methods when applied to small databases of 20 and 50 observations. In such
cases, OpRisk+ provides better results with the “round up” solution than with the “round” one. This
bias could be due to the fact that with the second method we tend to loose the “extreme value theory”
aspect of the model. Small databases tend indeed to lack extreme losses and taking the upper bound
limit value for the aggregation makes the resulting distribution’s tail fatter.
[TABLE 8]
4. Conclusions
This paper introduces a structural operational risk model, named OpRisk+, that has been inspired from
the well known credit risk model, CreditRisk+, which had characteristics transposable to the
operational risk modelling.
In a simulations setup, we work out aggregated loss distributions and operational Value-at-Risks
(OpVaR) corresponding to the confidence level required by Basel II. The performance of our model is
assessed by comparing our results to theoretical OpVaRs, to a lower bound issued from a simpler
approach, that is, the IMA approach of Alexander (2003), and to a more sophisticated approach using
two distributions in order to model the severity distributions of “normal” and “extreme” losses
proposed in Chapelle et al. (2008), or “CCHP” approach.
The results show that OpRisk+ produces higher OpVaRs than the lower bound of Alexander (2003),
but that it is not receptive enough to extreme events. On the other hand, our goal is not to produce an
adequate economic capital, but to try to propose a first solution to the lack of operational risk models.
Besides, whereas the CCHP approach has better sensitivity to very extreme losses, the simplicity of
OpRisk+ gives the model the advantage of requiring no large database in order to be implemented.
Specifically, we view the value-added of the OpRisk+ procedure as twofold. Firstly, it produces
average estimates of operational risk exposures that are very satisfactory at the 95% level, which
makes it a very useful complement to approaches that specifically target the extreme tail of the loss
18
distribution. Indeed, even though the performance of OpRisk+ is clearly not sufficient for the
measurement of unexpected operational losses as defined by the Basel II Accord (the VaR should be
measured with a 99.9% confidence level), it could be thought of as a sound basis for the measurement
of the body of losses; another more appropriate method must relay OpRisk+ for the measurement of
the far end of the distribution.
Secondly, despite the fact that we can not conclude that OpRisk+ is an adequate model to quantify the
economic capital associated to the bank’s operational risk in the LDA approach, its applicability to
approximate the loss distribution with small databases is proven. Even for such a small database as
one comprising 20 observations, the estimation could make it attractive as a complement to more
sophisticated approaches requiring large numbers of data per period. The fit is almost perfect when the
Pareto-distribution has a thin tail, and the OpVaR95 is the closest among the three specifications tested
when the tail gets fatter.
Of course, this approach is still subject to refinements, and could be improved in many ways. Indeed,
internal data rarely includes very extreme events (banks suffering those losses probably would no
more be there to tell us), whereas the last percentiles are very sensitive to the presence of those events.
The problem would therefore be to determine which weight to place on the internal data and on the
external ones. From our study, we could imagine that fitting a distribution calibrated with external
data or relying on EVT beyond the 95% percentile would justify the simultaneous use of OpRisk+
preferably to other models. This advantage can prove to be crucial for business lines or event types
where very few internal observations are available, and thus where most approaches such as the
CCHP would be powerless.
5. Notes
1 The Value-at-Risk (VaR) is the amount that losses will likely not exceed, within a predefined confidence level
and over a given time-period.
2 CreditRisk+’s authors argue that the exact amount of each loss cannot be critical in the determination of the
global risk.
3 See Appendix.
4 Several unit amounts were tested and did not lead to significantly different results. The size of the unit amount
should especially be chosen according to the number of losses: the bigger the number of losses in the database,
the bigger the unit amount, in order to lighten the computations.
5 More precisely the ALD is obtained through the n-fold convolution of the severity distribution with itself, n
being a random variable following the frequency density function.
19
6 While frequency could also be modelled with other discrete distributions such as the Negative Binomial for
instance, many authors use the Poisson assumption (see de Fontnouvelle, DeJesus-Rueff, Jordan and Rosengren,
2003, for instance).
7 This solution has been advocated by many other authors; see for instance King (2001), Cruz (2004),
Moscadelli (2004), de Fontnouvelle and Rosengren (2004) or Chavez-Demoulin, Embrechts and Neslehova
(2006).
8 See Embrechts, Klüppelberg and Mikosch (1997) for a comprehensive overview of EVT.
9 The unexpected loss is defined as the difference between the VaR at 99.9% and the expected loss.
10 Only 100 years of data were simulated for high-frequency databases as the computation becomes too heavy
for a too large number of data.
11 That is, every loss between 15000 and 25000 would be in band 20, instead of every loss between 10,000 and
20,000 being in band 20.
6. References
Alexander C. “Statistical Models of Operational Loss.” In: Operational Risk. Regulation, Analysis and
Management, Ft Prentice Hall Financial Times, 2003, pp129-170.
Barroin L, Ben Salem M. “Vers un risque opérationnel mieux géré et mieux contrôlé.» Banque
Stratégie 189, January 2002.
Basel Committee on Banking Supervision. “Basel II: International Convergence of Capital
Measurement and Capital Standards: A Revised Framework.” June 2004.
Basel Committee on Banking Supervision. “The 2002 Loss Data Collection Exercise for Operational
Risk: Summary of the Data Collected.” March 2003.
Basel Committee on Banking Supervision. “Overview of the New Basel Capital Accord”,
Consultative Document, April 2003.
Basel Committee on Banking Supervision. “Operational Risk Data Collection Exercise.» June 2002.
Basel Committee on Banking Supervision, “Operational Risk”, Supporting Document to the New
Basel Capital Accord, January 2001.
Basel Committee on Banking Supervision, “Working Paper on the Regulatory Treatment of
Operational Risk.» September 2001.
20
Borysiak M. Geginat M. Herkes R. “Operational Risk: How to Create and Use Models in the Context
of the Advanced Measurement Approach AMA?” Le Mensuel d’Agefi Luxembourg,
http://www.agefi.lu, Edition of April 2003, Article N°5343.
Chapelle, A. Crama, Y. Hübner, G. Peters, J. “Practical Methods for Measuring and Managing
Operational Risk in the Financial Sector: A Clinical Study.” Journal of Banking & Finance 32
(2008), 1049–1061.
Chapelle, A. Hübner, G. Peters, J.- “Le risque opérationnel : Implications de l’Accord de Bâle pour le
secteur financier.” Cahiers Financiers, Larcier, Brussels, 2004.
Credit Suisse First Boston. “Creditrisk+ : A Credit Risk Management Framework.» Credit Suisse
Financial Products, 1997.
Chavez-Demoulin, V. Embrechts, Neslehova, J. “Quantitative Models for Operational Risk:
Extremes, Dependence and Aggregation.» Journal of Banking and Finance 30, 2635-2658.
Crouhy M. Galai D. Mark R. “A Comparative Analysis of Current Credit Risk Models”, Journal of
Banking and Finance 24 (2000), 59-117.
Cruz, M. Operational Risk Modelling and Analysis: Theory and Practice, Risk Waters Group,
London, 2004.
De Fontnouvelle Rosengren E. “Implications of Alternative Operational Risk Modeling Techniques.»
Federal Reserve Bank of Boston: http://www.bos.frb.org/bankinfo/qau/pderjj604.pdf., 2004.
De Fontnouvelle Dejesus-Rueff, V. Jordan, J. Rosengren E. “Capital and Risk: New Evidence on
Implications of Large Operational Losses.” Federal Reserve Bank of Boston, Working Paper,
September 2003.
De Fontnouvelle et al. “Using Loss Data to Quantify Operation Risk.» Working Paper, April 2003.
Degen M., Embrechts P., Lambrigger D. “The quantitative modelling of operational risk: between g-
and-h and EVT.” Astin Bulletin 37, Vol 2. (2007), 265-291.
Embrechts P., Kluppelberg C., Mikosch T. “Modelling Extremal Events for Insurance and Finance”,
Springer-Verlag, Berlin, 1997.
Finger C. “Sticks and Stones”, Riskmetrics Grou Working Paper, 1999.
21
Frachot A. Georges Roncalli T. “Loss Distribution Approach for Operational Risk.” Working Paper,
Groupe de Recherche Opérationnelle, Crédit Lyonnais, France, 2001.
Frachot A. Roncalli T. “Mixing Internal and External Data for Managing Operational Risk.” Groupe
de Recherche Opérationnelle, Crédit Lyonnais, France, January 2002.
Gordy M. “A Comparative Anatomy of Credit Risk Models.” Journal of Banking and Finance 24
(2000), 119-149.
Kern M. Rudolph B. “Comparative Analysis of Alternative Credit Risk Models – An Application on
German Middle Market Loan Portfolios.” Center for Financial Studies, No 2001/3.
King, J. “Measurement and Modelling” Operational Risk, Wiley, New York, 2001.
Koyluoglu, H. Hickman A. “A Generalized Framework for Credit Risk Portfolio Models”,
Http://Www.Erisk.Com, 1998.
Lopez J.A. Saidenberg M.R. “Evaluating Credit Risk Models.” Journal of Banking and Finance 24
(2000), 151-165.
Merton R. “On the Pricing of Corporate Debt: the Risk Structure of Interest Rates.” Journal of Finance
28 (1974), 449-470.
Mori T. Harada E. “Internal Risk Based Approach - Evolutionary Approaches to Regulatory Capital
Charge for Operational Risk.” Bank of Japan, Working Paper, 2000.
Mori T. Harada E. “Internal Measurement Approach to Operational Risk Capital Charge.” Bank of
Japan, Discussion Paper, 2001.
Moscadelli M. “The Modelling of Operational Risk: Experience with the Analysis of the Data
Collected by the Basel Committee.” Banca d’Italia 517 (2004).
Sahay.A., Wan Z., Keller B. “Operational Risk Capital: Asymptotics in the case of Heavy-Tailed
Severity.” Journal of Operational Risk, Vol. 2 No.2 (2007).
Wilson R. “Corporate Senior Securities: Analysis and Evaluation of Bonds, Convertibles and
Preferred.” Chicago: Probus, 1987.
Wilson R. “Portfolio Credit Risk (I, Ii)”, Risk Magazine, September/October 1997, pp. 111-117 and
pp. 56-61.
22
7. Tables
Loss Amount Loss in L round-off loss band j
(LGE) νj
1 500 1.5 2.00 2
2 508 2.51 3.00 3
3 639 3.64 4.00 4
1 000 1.00 1.00 1
1 835 1.84 2.00 2
2 446 2.45 3.00 3
7 260 7.26 8.00 8
Table 1 - Allocating losses to bands.
νj μj εj
1 9 9
2 121 242
3 78 234
4 27 108
5 17 85
6 15 90
7 8 56
8 4 32
Table 2 - Exposure, number of events and expected loss.
100 50 40 30 20 10 8 6
VaR99.9 131.81 72.75 60.45 47.81 34.71 20.66 17.63 14.45
3.18 3.22 3.23 3.25 3.29 3.37 3.41 3.45
0.32 0.46 0.51 0.59 0.74 1.07 1.21 1.41
5 4 3 2 1 0.9 0.8 0.7
VaR99.9 12.77 10.96 9.13 7.11 4.87 4.55 4.23 3.91
3.48 3.48 3.54 3.62 3.87 3.85 3.84 3.84
1.55 1.74 2.04 2.56 3.87 4.06 4.29 4.59
0.6 0.5 0.4 0.3 0.2 0.1 0.05 0.01
VaR99.9 3.58 3.26 2.91 2.49 2.07 1.42 1.07 0.90
3.85 3.90 3.97 4.00 4.19 4.17 4.54 8.94
4.97 5.51 6.27 7.30 9.36 13.21 20.31 89.40
Table 3 - Gamma and phi values (no loss severity variability) (source: Alexander (2003), p151).
23
Panel A : Thin-tailed-Pareto distribution (shape parameter = 0.3)
Poisson parameter μ
Number of losses.
20 50 200 300
Theoretical Mean 2,857 7,143 28,571 42,857
Mean 2,845 7,134 28,381 42,886
Standard deviation 287 472 847 1,118
Median 2,796 7,078 28,172 42,763
Maximum 4,683 9,026 30,766 45,582
Minimum 2,268 6,071 26,713 40,383
Number of simulated years 1000 1000 100 100
Panel B : Medium-tailed-Pareto distribution (shape parameter = 0.5)
Poisson parameter μ
Number of losses
20 50 200 300
Theoretical Mean 4,000 10,000 40,000 60,000
Mean 3,924 9,913 39,871 59,431
Standard deviation 1,093 1,827 3,585 5,504
Median 3,676 9,594 39,777 57,947
Maximum 15,680 29,029 54,242 91,182
Minimum 2,567 7,097 33,428 52,436
Number of simulated years 1000 1000 100 100
Panel C : Fat-tailed-Pareto distribution (shape parameter = 0.7)
Poisson parameter μ
Number of losses.
20 50 200 300
Theoretical Mean 6,667 16,667 66,667 100,000
Mean 6,264 16,165 61,711 93,724
Standard deviation 5,940 13,018 13,899 24,514
Median 5,180 13,721 57,713 87,646
Maximum 157,134 265,621 137,699 248,526
Minimum 2,646 8,304 45,315 69,991
Number of simulated years 1000 1000 100 100
Table 4 - Characteristics of Twelve Databases under OpRisk+.
24
Panel A : Thin-tailed-Pareto distribution (shape parameter = 0.3)
N = 20 N = 50
Simulated
OpRisk+ Simulated
OpRisk+
Mean Δ S.D. Min Max Mean Δ S.D. Min Max
OpVaR90 3770 3880 3% 448 3080 7580 8573 8882 4% 706 7530 16980
OpVaR95 4073 4173 2% 505 3290 8020 9030 9334 3% 769 7880 17010
OpVaR99 4712 4744 1% 612 3710 8110 9942 10209 3% 906 8560 17020
OpVaR99,9 5596 5410 -3% 717 3780 9800 11141 11250 1% 1241 9340 30010
Panel B : Medium-tailed-Pareto distribution (shape parameter = 0.5)
N = 20 N = 50
Simulated
OpRisk+ Simulated
OpRisk+
Mean Δ S.D. Min Max Mean Δ S.D. Min Max
OpVaR90 5579 5672 2% 2209 3470 29360 12630 12855 2% 3102 8760 51800
OpVaR95 6364 6247 -2% 2896 3720 40860 13862 13734 -1% 3838 9190 70420
OpVaR99 8966 7329 -18% 3940 4190 53610 18051 15410 -15% 4900 10020 91180
OpVaR99,9 18567 8626 -54% 5120 4750 66930 33554 17338 -48% 6013 10990 112940
Panel C : Fat-tailed-Pareto distribution (shape parameter = 0.7)
N = 20 N = 50
Simulated
OpRisk+ Simulated
OpRisk+
Mean Δ S.D. Min Max Mean Δ S.D. Min Max
OpVaR90 9700 11410 18% 10214 4300 106700 22495 23992 7% 25901 11700 526900
OpVaR95 12640 12931 2% 12441 4600 142450 28103 27089 -4% 37495 12300 777800
OpVaR99 27261 15583 -43% 15736 5150 189050 55994 32020 -43% 49854 13450 1033000
OpVaR99,9 114563 18726 -84% 19524 5850 236200 220650 38761 -82% 69020 14750 1290300
Table 5 - Values-at-Risk generated by OpRisk for small databases, with 20 and 50 loss events.
25
Panel A : Thin-tailed-Pareto distribution (shape parameter = 0.3)
N = 200 N = 300
Simulated
OpRisk+ Simulated
OpRisk+
Mean Δ S.D. Min Max Mean Δ S.D. Min Max
OpVaR90 31448 33853 7% 14819 31780 36640 46355 56470 22% 1240 53950 59800
OpVaR95 32309 34728 7% 15435 32600 37780 47403 57683 22% 1305 55100 61200
OpVaR99 33995 36397 7% 16677 34180 40340 49420 59992 21% 1431 57200 63950
OpVaR99,9 36063 38310 6% 18197 36000 43740 51750 62628 21% 1589 59650 67150
Panel B : Medium-tailed-Pareto distribution (shape parameter = 0.5)
N = 200 N = 300
Simulated
OpRisk+ Simulated
OpRisk+
Mean Δ S.D. Min Max Mean Δ S.D. Min Max
OpVaR90 45757 51836 13% 5875 43500 79650 67104 75723 13% 9683 66600 146300
OpVaR95 48259 53816 12% 6935 44650 89600 70264 78161 11% 11515 68100 164700
OpVaR99 55919 57668 3% 8940 46900 105300 79718 82817 4% 14571 70950 193300
OpVaR99,9 83292 62237 -25% 11431 49450 123600 113560 88309 -22% 18259 74250 226700
Panel C : Fat-tailed-Pareto distribution (shape parameter = 0.7)
N = 200 N = 300
Simulated
OpRisk+ Simulated
OpRisk+
Mean Δ S.D. Min Max Mean Δ S.D. Min Max
OpVaR90 82381 82539 0% 24959 57200 218600 120654 119943 -1% 34805 86100 364800
OpVaR95 96971 88248 -9% 30247 58950 247000 139470 127037 -9% 42644 88400 436550
OpVaR99 166962 98972 -41% 39404 62300 303400 234442 55846 -40% 55846 92850 550800
OpVaR99,9 543597 111875 -79% 50432 66150 373100 733862 156642 -79% 71810 98000 687000
Table 6 - OpVaRs generated by OpRisk+ for databases with 200 and 300 loss events.
26
Panel A : Thin-tailed-Pareto distribution (shape parameter = 0.3)
N = 200 N = 300
Simulated
OpRisk+ Simulated
OpRisk+
Roundup Δ Round Δ Roundup Δ Round Δ
OpVaR90 31448 33853 8% 30576 -3% 46355 56470 22% 43558 -6%
OpVaR95 32309 34728 7% 31404 -3% 47403 57683 22% 44563 -6%
OpVaR99 33995 36397 7% 32991 -3% 49420 59992 21% 46486 -6%
OpVaR99,9 36063 38310 6% 34813 -3% 51750 62628 21% 48687 -6%
Panel B : Medium-tailed-Pareto distribution (shape parameter = 0.5)
N = 200 N = 300
Simulated
OpRisk+ Simulated
OpRisk+
Roundup Δ Round Δ Roundup Δ Round Δ
OpVaR90 45757 51836 13% 44338 -3% 67104 75723 13% 64523 -4%
OpVaR95 48259 53816 12% 46222 -4% 70264 78161 11% 66849 -5%
OpVaR99 55919 57668 3% 49885 -11% 79718 82817 4% 71296 -11%
OpVaR99,9 83292 62237 -25% 54257 -35% 113560 88309 -22% 76544 -33%
Panel C : Fat-tailed-Pareto distribution (shape parameter = 0.7)
N = 200 N = 300
Simulated
OpRisk+ Simulated
OpRisk+
Roundup Δ Round Δ Roundup Δ Round Δ
OpVaR90 82381 82539 0% 75696 -8% 120654 119943 -1% 112596 -7%
OpVaR95 96971 88248 -9% 81375 -16% 139470 127037 -9% 120850 -13%
OpVaR99 166962 98972 -41% 91991 -45% 234442 55846 -76% 135481 -42%
OpVaR99,9 543597 111875 -79% 104699 -81% 733862 156642 -79% 152904 -79%
Table 7 - Comparison of OpRisk+ with an upper bound limit value (rounded up) and an average
value (rounded) for each band, for large databases.
27
Panel A : Thin-tailed-Pareto distribution (shape parameter = 0.3)
N = 20 N = 50
Simulated
OpRisk+ Simulated
OpRisk+
Roundup Δ Round Δ Roundup Δ Round Δ
OpVaR90 3770 3880 3% 3535 -6% 8573 8882 4% 8074 -6%
OpVaR95 4073 4173 2% 3815 -6% 9030 9334 3% 8501 -6%
OpVaR99 4712 4744 1% 4363 -7% 9942 10209 3% 9332 -6%
OpVaR99,9 5596 5410 -3% 5010 -10% 11141 11250 1% 10311 -7%
Panel B : Medium-tailed-Pareto distribution (shape parameter = 0.5)
N = 20 N = 50
Simulated
OpRisk+ Simulated
OpRisk+
Roundup Δ Round Δ Roundup Δ Round Δ
OpVaR90 5579 5672 2% 5332 -4% 12630 12855 2% 11323 -10%
OpVaR95 6364 6247 -2% 5901 -7% 13862 13734 -1% 12152 -12%
OpVaR99 8966 7329 -18% 6945 -23% 18051 15410 -15% 13668 -24%
OpVaR99,9 18567 8626 -54% 7904 -57% 33554 17338 -48% 14377 -57%
Panel C : Fat-tailed-Pareto distribution (shape parameter = 0.7)
N = 20 N = 50
Simulated
OpRisk+ Simulated
OpRisk+
Roundup Δ Round Δ Roundup Δ Round Δ
OpVaR90 9700 11410 18% 9413 -3% 22495 23992 7% 25235 12%
OpVaR95 12640 12931 2% 10914 -14% 28103 27089 -4% 28537 2%
OpVaR99 27261 15583 -43% 13353 -51% 55994 32020 -43% 33837 -40%
OpVaR99,9 114563 18726 -84% 16290 -86% 220650 38761 -82% 40024 -82%
Table 8 - Comparison of OpRisk+ with an upper bound limit value (round up) and an average
value (round) for each band, for small databases.
28
8. Figures
Figure 1. Aggregated loss distribution for a series of 200 loss events characterized by a
Pareto(100;0.3).
PANEL A : THIN-TAILED DISTRIBUTION (TAIL INDEX = 0,3)
0
0,001
0,002
0,003
0,004
0,005
0,006
0,007
0,008
0,009
24900 27400 29900 32400 34900 37400 39900 42400 44900 47400 49900 52400 54900
Aggregated loss
Pro
ba
bil
ity
3000
3500
4000
4500
5000
5500
6000
6500
7000
7500
0,9 0,95 0,99 0,999
OpVaRs for N = 20
7000
8000
9000
10000
11000
12000
13000
0,9 0,95 0,99 0,999
OpVaRs for N = 50
30000
31000
32000
33000
34000
35000
36000
37000
38000
39000
40000
0,9 0,95 0,99 0,999
OpVaRs for N = 200
45000
47000
49000
51000
53000
55000
57000
59000
61000
63000
65000
0,9 0,95 0,99 0,999
OpVaRs for N = 300
29
PANEL B : THIN-TAILED DISTRIBUTION (TAIL INDEX = 0,5)
PANEL C : THIN-TAILED DISTRIBUTION (TAIL INDEX = 0,7)
Figure 2 - Comparison of CCHP, OpRisk+ and the lower bound proposed by Alexander (2003).
4000
6000
8000
10000
12000
14000
16000
18000
20000
0,9 0,95 0,99 0,999
OpVaRs for N = 20
10000
15000
20000
25000
30000
35000
0,9 0,95 0,99 0,999
OpVaRs for N = 50
40000
45000
50000
55000
60000
65000
70000
75000
80000
85000
90000
0,9 0,95 0,99 0,999
OpVaRs for N = 200
60000
70000
80000
90000
100000
110000
120000
0,9 0,95 0,99 0,999
OpVaRs for N = 300
0
20000
40000
60000
80000
100000
120000
140000
0,9 0,95 0,99 0,999
OpVaRs for N = 20
0
50000
100000
150000
200000
250000
0,9 0,95 0,99 0,999
OpVaRs for N = 50
0
100000
200000
300000
400000
500000
600000
0,9 0,95 0,99 0,999
OpVaRs for N = 200
0
100000
200000
300000
400000
500000
600000
700000
800000
0,9 0,95 0,99 0,999
OpVaRs for N = 300
30
Figure 3. OpVaR levels generated by CCHP, OpRisk+ and theoretical OpVaR for databases of
20 events characterized by a Pareto distribution with parameters (100;0.3), (100;0.5) and
(100.0.7), namely 20_3, 20_5 and 20_7, respectively, and for databases of 50 events
characterized by the same distribution, that is respectively 200_3, 200_5 and 200_7.
20_3
3.000
4.000
5.000
6.000
7.000
8.000
OpVaR_90 OpVaR_95 OpVaR_99 OpVaR_99.9
Theory CCHP OpRisk+
200_3
30.000
32.000
34.000
36.000
38.000
40.000
OpVaR_90 OpVaR_95 OpVaR_99 OpVaR_99.9
Theory CCHP OpRisk+
20_5
5.000
10.000
15.000
20.000
OpVaR_90 OpVaR_95 OpVaR_99 OpVaR_99.9
Theory CCHP OpRisk+
20_7
0
20.000
40.000
60.000
80.000
100.000
120.000
140.000
OpVaR_90 OpVaR_95 OpVaR_99 OpVaR_99.9
Theory CCHP OpRisk+
200_7
0
100.000
200.000
300.000
400.000
500.000
600.000
OpVaR_90 OpVaR_95 OpVaR_99 OpVaR_99.9
Theory CCHP OpRisk+
200_5
40.000
50.000
60.000
70.000
80.000
90.000
OpVaR_90 OpVaR_95 OpVaR_99 OpVaR_99.9 Theory CCHP OpRisk+
31
9. Appendix - CreditRisk+: The distribution of Default losses Calculation
procedurei
CreditRisk+ mathematically describes the random effect of the severity distribution through its
probability generating function G(Z):
Comparing this definition with the Taylor series expansion for G(z), the probability of a loss of n x L,
An, is given by:
In CreditRisk+, G(Z) is given in closed form by :
Therefore, using Leibniz formula we have:
However
and by definition
Therefore
Using the relation jjj μ.νε, the following recurrence relationship is obtained:
i Source : Credit Suisse (1997); “CreditRisk+ : A Credit Risk Management Framework”, Credit Suisse Financial Products, Appendix A, Section A4, p36.